David Grijalva

AI products, designed and built end to end.

I turn complex ideas into dependable products by connecting strategy, experience design, architecture, implementation, and deployment.

View selected work About me

From product thesis to operational proof

This portfolio connects product strategy with hands-on execution across agentic software, creative technology, conversation intelligence, and production infrastructure.

Products and systems built around difficult constraints

Houdini 22 MCP

A production-minded control layer for operating SideFX Houdini 22 through typed, observable, and recoverable agent actions.

  • 339 typed MCP tools spanning OBJ, SOP, DOP, CHOP, Copernicus, ROP, Solaris/USD, TOP/PDG, VOP/MAT, APEX, animation, HDAs, and scene operations.
  • Live-verified on macOS Apple Silicon and Windows with Houdini 22.0.368.
  • Dry-run graph preflight, atomic multi-node construction, rollback, cook verification, durable background work, and retained results.
Read the case study
Live control pathSession connected
01Agent clientMCP requests
02External serverTyped tools + policy
03Local bridgeAuthenticated + bounded
04Houdini 22Main-thread scene control
node.createdparm.changedinput.rewiredcook.verifiedjob.retained

VectorCI

A capability-aware control plane for repository-isolated, self-hosted CI execution with explicit proof of what each route can actually do.

  • 81.8% estimated average cost reduction across two practical hosted-CI baselines: GitHub Actions after its included allowance and my typical paid CircleCI month.
  • Signed GitHub workflow-job event handling with persisted deliveries, jobs, and one-capacity pool slots.
  • Repository-bound ephemeral Railway Linux workers with measured execution and cost evidence.
Read the case study
Repository-isolated lifecycleCapacity 1 / 1
  1. 01Job queued
  2. 02Signed event
  3. 03Capability matched
  4. 04Runner starts
  5. 05Job executes
  6. 06Evidence recorded
  7. 07Runner stops
route: linux-generalisolation: repositoryfallback: noneteardown: required

VectorVoice

A privacy-first conversation workspace that converts temporary speech into source-linked decisions, actions, and commitments.

  • Bot-free capture, audio/video intake, provider-neutral transcription, and a separately deployable conversation-intelligence service.
  • Organization-scoped meetings, notes, source evidence, decisions, actions, and authenticated workspace question answering.
  • Verified temporary-audio deletion, expiry sweeps, full-transcript deletion, and retained cited evidence moments.
Read the case study
VectorVoice meeting evidence view showing cited moments and retained decisions

NeuroTopo

An artist-controlled Houdini retopology system with atomic editing, recoverable state, and a native topology-intelligence architecture.

  • A working Houdini Quad Draw-style viewer workflow for points, quads, relax, extend, tweak, split, tube creation, selection, cleanup, symmetry, suggestions, and validation.
  • Atomic and recoverable topology operations with stable IDs, candidate isolation, compare-and-swap publication, local legality, and hash-chained portable state.
  • Native persistent surface and topology contexts, affected-region updates, hardware-aware scheduling, and retained overlay paths.
Read the case study
NeuroTopo retopology viewer state running in a Houdini viewport

Capabilities behind the work

Technical depth applied to product decisions and operating risk

01

Agentic product design

Translate open-ended requests into bounded actions with visible state, approval, and recovery.

02

Product strategy and architecture

Turn a product thesis into clear responsibilities, durable contracts, and an executable release plan.

03

Creative technology

Build artist-facing software where real-time interaction, native performance, and creative control must coexist.

04

Privacy and trust

Make authorization, evidence, retention, approval, and recovery part of the product experience.

05

Production infrastructure

Connect deployment claims to exact revisions, environments, execution routes, and runtime evidence.

06

Operational clarity

Define what the system owns, how it fails, and what evidence supports each release decision.

Systems in development and validation

Vector Code

Before Codex and Claude added mature IDE and mobile workflows, I preferred running agents in the terminal—but reviewing changed files meant using Vim or opening a separate IDE. I built Vector Code to bring terminal-first agent work and structured code review into one project-scoped workbench.

Principal challengeKeep files, editor tabs, terminals, source control, tasks, and agent sessions aligned to the active project, then pair the desktop through an encrypted mobile relay so I could leave my laptop, check progress, and give new direction from my phone.

Private repository

Google Multi-Account Connectors

A hosted plugin and MCP service that keeps multiple Gmail, Calendar, and Drive identities available through one authenticated account set.

Principal challengeSeparate OpenAI identity from Google authorization, encrypt every refresh token per user, and preserve the correct account context securely across devices.

Private repository

LiquidSolver

A native Houdini SOP plugin for a CUDA-only water solver with strict failure behavior and sparse large-body foundations.

Principal challengeKeep simulation, cache, packaging, GPU support, and failure claims bound to the exact tested environment while building toward broader production-scale use.

Private repository

VectorVault

A client-encrypted password-manager foundation spanning a web vault, browser extension, ciphertext-only sync, and native AutoFill adapters.

Principal challengePreserve cryptographic and platform trust boundaries across web and native clients before independent review, penetration testing, interoperability, and signed releases.

Private repository

Vector + VectorPlatform

An AI-native company context and execution product backed by reusable, product-neutral connector contracts and infrastructure.

Principal challengeSeparate shared mechanics from product-owned tenancy, credentials, and proof—and distinguish a 39-provider contract catalog from the smaller set verified live in staging and production.

Private repository

Polaris Finder

After a memorable Polaris flight to Iceland, I built a flexible-date finder for discovering another destination where the cash-and-miles mix makes the experience attainable.

Principal challengeThe playful idea became a serious browser-automation problem: search flexible dates, verify exact United Polaris segments and live fares, and handle sign-in, MFA, CAPTCHA, rate limits, and airline outages without returning stale or misleading results.

Private repository